1. 3.3. Case-Based Reasoning (CBR)
Knowledge Component 3: Acquiring Data and Knowledge
3.3. Case-Based Reasoning (CBR)
2nd Edition
Ian F. C. Smith
EPFL, Switzerland

2. Module Information Novice Intended audience Key words
Case-based reasoning, information
retrieval, adaptation, case storage
Author
Ian F.C. Smith, EPFL, Switzerland 2

3. What there is to learn
At the end of this module, there will be answers to the following questions:
Why use cases?
How can engineers ensure that good results are obtained?
What are the long-term challenges? 3

4. Steps in Case-Based Reasoning
Outline
Introduction
Steps in Case-Based Reasoning
CBR Cycle
CBR in Everyday Life

5. Principle
Case-based reasoning (CBR) involves finding solutions to new tasks by reusing good solutions to old tasks.
CBR systems have many advantages over other types of computer systems for decision support.
Some of them are given in the following slides.

6. Advantages of CBR
The process of storing cases often involves transformations that are computationally less expensive than those necessary with other representations.
A good case is often an easy shortcut in the search for good solutions.
The closed-world assumption associated with abductive tasks (Module 1.3) is explicitly related to the number of available cases.

7. Advantages (cont'd.)
When only small changes are made to the old case, intrinsic advantages of the old case are transferred to the new task.
Cases are often the best way to represent knowledge, especially when the available models are unreliable.
The capability of the system can be enhanced by simply adding a case.

8. Advantages (cont'd.)
Large case-based reasoning systems are currently used for diagnosis, maintenance, planning and customer support.
There is much potential in civil engineering.

9. Steps in Case-Based Reasoning
Outline
Introduction
Steps in Case-Based Reasoning
CBR Cycle
CBR in Everyday Life

10. Five Steps
There are five main steps in the development and use of a CBR system.
Representation
Retrieval
Adaptation
Storage
Maintenance

11. Step 1: Representation
This step involves decisions related to how the case is stored in the computer.
For example, if cases are to be stored in a relational database, appropriate attributes have to be selected.
This is an important step since it can greatly influence the quality of support provided.
One has to know clearly for what purpose the cases are intended before finalizing the representation.

12. Step 1: Representation (cont'd.)
Characteristics of both task and solution must be represented.
Task characteristics are used for case retrieval (Step 2) and solution characteristics are used for case adaptation (Step 3).
Case =
Task characteristics
Solution characteristics +

13. Step 2: Retrieval
This step involves the selection of a case (or cases) that have task characteristics which are similar to the new task.
Here, the challenge involves defining what the word “similar” means.
Hundreds of similarity metrics exist. The most common employ “nearest neighbor” methods.

14. Steps 2: Retrieval (cont'd.)
Below is a simple example of a nearest neighbor method for numerical attributes.
For a simple nearest neighbor calculation:
For each task attribute, define a maximum value for the difference, Max.diff. This value sets the limit for similarity. Beyond this value, two attribute values are not similar.

15. Steps 2: Retrieval (cont'd.)
Define weight factors wf using a scale that describes the relative importance of each task attribute. An iterative loop begins here:
For the given case, for each task attribute, calculate the difference (absolute value) between the value for the attribute in the case and in the new task. Divide this difference by Max.diff. Values greater than 1 are set to 1. This result is called Distance.

16. Steps 2: Retrieval (cont'd.) Similarity = 1-Distance
Evaluate the similarity variable for each task attribute
Similarity = 1-Distance
Calculate a weighted overall similarity measure between two cases [0…1]
and return to step 3.

17. Steps 2: Retrieval (cont'd.)
Report all cases having values for overall similarity that are above a pre-set level.
If possible, it is always preferable to report more than one case. Engineers often want computers to give them choices, not answers, see Module 1.3.
At this point, users may choose to ignore certain cases according to considerations that are not represented in the task attributes.

18. Steps 3: Adaptation
The adaptation stage is the most difficult to implement on a computer.
Here, the solution attributes of the retrieved case(s) are modified to suit the new task. This is seldom automated.
User interaction is often essential in order to bring in the domain knowledge necessary for creating a useful solution.

19. Steps 3: Adaptation (cont'd.)
While this stage is poorly supported in many commercial systems, it is the crucial stage of case-based reasoning.
Without this stage, case-based reasoning would be no more than an exercise in information retrieval.

20. Steps 4: Storage
The storage stage involves storing the newly found solution in the case base for future use.
If the solution is within the scope of solutions that were envisaged when case attributes were defined, this stage is straightforward.

21. Steps 5: Maintenance
The maintenance stage involves elimination of cases that are no longer relevant to the tasks that will come in the future.
This stage may also include the task of ensuring that the cases uniformly cover the entire range of possible attribute values.

22. Steps in Case-Based Reasoning
Outline
Introduction
Steps in Case-Based Reasoning
CBR Cycle
CBR in Everyday Life

23. The CBR Cycle Initial Cases Storage Solution Data Base Adaptation
Maintenance
Retrieval
New Task

24. Steps in Case-Based Reasoning
Outline
Introduction
Steps in Case-Based Reasoning
CBR Cycle
CBR in Everyday Life

25. CBR in Everyday Life
The concept of CBR is used in other fields where decisions are required in “open worlds”.
Medicine
Doctors remember previous medical cases, especially for rare symptoms
Law
Case histories are always consulted
Judgments often depend on precedents
Management
Learning through study of previous cases

26. Suitability of CBR Well suited tasks Classification Diagnosis
Prediction
Control
Harder for CBR
Synthesis of systems
Engineering design
Planning of complex projects

27. Review Quiz List the process steps for all CBR methods.
What are characteristics that indicate that a CBR approach may be suitable?
What is the end criterion for retrieval?

28. Answers to Review Quiz List the process steps for all CBR methods.
Retrieve the most similar case(s) comparing the present case to the library of past cases
Select a retrieved case to solve the current task
Adapt the old case to make a new solution
Retain the new solution as a new case
Maintain the case base

29. Answers to Review Quiz
What are characteristics that indicate that a CBR approach may be suitable?
Records of previously solved problems exist
Remembering previous experience is useful
Cases can be represented and stored without excessive use of resources
Adaptation of cases is feasible
Maintenance that ensures complete and uniform coverage is possible

30. Answers to Review Quiz What is the end criterion for retrieval?
Retrieval is complete when the best matching case is found.
Subtasks involve:
identifying a set of relevant problem descriptors
matching the case and returning a set of sufficiently similar cases (given a similarity criterion)
selecting the best case from the set of cases returned

31. Summary
CBR involves finding solutions to new tasks by reusing good solutions to old tasks.
The capability of a system can be increased by simply adding a case.
There are 5 stages in the CBR methodology:
Representation
Retrieval
Adaptation
Storage
Maintenance

32. Answers to initial questions
Why use cases?
Cases provide good answers to previous challenges. They are shortcuts to solutions; they may contain valuable implicit knowledge; they are useful when no domain models exist; adding knowledge involves just adding a case.
How can engineers ensure that good results are obtained?
Cases need to uniformly cover all possible applications
What are the long-term challenges?
Maintenance: dropping irrelevant cases and ensuring uniform coverage.

33. Further Reading
J. Kolodner. Case-Based Reasoning. Morgan Kaufmann Publishers, San Mateo, CA, 1993
Raphael, B. and Smith, I.F.C. Fundamentals of Computer-Aided Engineering, Wiley, 2003
Case-Based Reasoning on the Web, University of Kaiserslautern,